search

ryokubozono / deepstream-rtspsrc-yolo

4 stars 1 forks source link

readme

Deepstream rtspsrc YOLO

system_drawing

This app detect objects from RTSP source and create RTSP output.

This app is made from Deepstream sample-apps.

YOLOv4 pre-trained model is trained using https://github.com/AlexeyAB/darknet.

ONNX model is created using https://github.com/Tianxiaomo/pytorch-YOLOv4.

Pre-trained model detects 10 classes of shells.

When RTSP source is disconnected, the APP wait for restarting RTSP source and try reset pipeline.

Demo

demo

Environment

Jetson

Set up Jetson

Doc for Jetson Nano: https://developer.nvidia.com/embedded/learn/get-started-jetson-nano-devkit

YOLO

Build custom YOLOv4

cd yolo/nvdsinfer_custom_impl_Yolo
make

You will find yolo/nvdsinfer_custom_impl_Yolo/libnvdsinfer_custom_impl_Yolo.so and use it as pgie's custom-lib-path.

Download pre-trained ONNX model

This YOLOv4 model detects 10 classes of shell.

Google Drive: https://drive.google.com/file/d/1MDd2u1Rn9xemOjd4fjJhm8SIHGqoieF8/view?usp=sharing

Convert ONNX model to TensorRT model

Set path to trtexec.

export PATH=/usr/local/cuda-10.2/bin:/usr/src/tensorrt/bin:$PATH

Convert <onnx_file_name> to <engine_file_name>.

cd yolo
trtexec \
    --onnx=<onnx_file_name> \
    --explicitBatch \
    --saveEngine=<engine_file_name> \
    --minShapes=input:1x3x416x416 \
    --optShapes=input:1x3x416x416 \
    --maxShapes=input:1x3x416x416 \
    --shapes=input:1x3x416x416 \
    --fp16

Run prediction

Compile app.

export CUDA_VER=10.2
make

Run app with <rtsp_source_uri>.

./deepstream-rtspsrc-yolo-app \
    <rtsp_source_uri>

Rum app with multi streams.

./deepstream-rtspsrc-yolo-app \
    <rtsp_source_uri_1> \
    <rtsp_source_uri_2>

Watch predicted streams.

rtsp://<Jestson's IP>:8554/dt-test

Others

PGIE's interval

Update interval in config/pgie_config.txt to decrease Jetson's work load.

This config means skip frames of RTSP source every interval number.

interval=5

Custom YOLO model

If you want to use custom YOLOv4 or YOLOv4-tiny model, you need the following tasks.

Custom classes number

If you want to change pgie classes number form 10, you neet to change